Classification of Glioma by Exploring Wavelet-based Radiomic Features and Machine Learning Techniques using Brats Dataset

The unusual growth of brain cells leads to brain tumors, out of which gliomas are an aggressive type with increased mortality rate and having different grading. Classifying gliomas into high-grade and low-grade gliomas helps oncologists better manage patient databases and increase the efficiency of treatment plans. Using radiomic features and machine learning techniques on MRI brain images for glioma grading is an emerging research topic. Recently the research on brain tumor MRI images has been conducted on the Brats dataset, which contains training and validation MRI brain images of different patients. In this work, an attempt is made to extract the most relevant features of MR images using pyradiomics and wavelet filters which is helpful in various image mining techniques to improve the overall efficiency. The relevance of extracted features is experimented with using various classifiers, including Support vector machine, random forest, Naive Bayes, Decision Trees, Bagging classifier, K Nearest Neighbor, and logistic regression. The accuracy and sensitivity of classifiers and the effect of feature reduction techniques are analyzed and compared with the identified features. The experimental results show that the Support Vector machine coupled with the factor analysis reduction technique outperforms other classifiers used in this study in terms of stability (RSD) and mean accuracy of about 97%.